feat(data_process): 模型缓存统一仓库根 .cache 并修复 PDF 页眉页脚清理

- 新增 app/core/cache_paths.py:HF_HOME / tiktoken 缓存统一指向 <repo>/.cache,
  本地与 Docker 路径一致,离线部署打包 .cache 即可
- Dockerfile.backend 的 tiktoken 词表改用官方 SHA 文件名,避免运行时回退重建
- 修复 layout_hybrid 路径不运行 detect_pdf_document_noise 的缺陷:
  needs_pdf_noise 不再与 needs_layout_raw 互斥,PDF 智能预处理在版面切分下也生效
- 新增 layout_noise.py:识别跨页重复的页眉表格标签组并按行剔除,
  解决 docling layout 模型把中文企业 PDF 页眉识别成普通 Table 导致清不掉的问题
- 回收 HybridChunker 丢弃的末尾孤立标题,找回章节标题内容
This commit is contained in:
caoxiaozhu
2026-08-21 10:15:08 +08:00
parent 3b9361c237
commit 03254f8196
14 changed files with 479 additions and 32 deletions

View File

@@ -1686,7 +1686,6 @@ def _prepare_preview_items(
)
needs_pdf_noise = (
is_unstructured
and not needs_layout_raw
and source_format == "pdf"
and bool(
preprocess_options & {"clean_invalid", "clean_invalid_content"}
@@ -1720,17 +1719,18 @@ def _prepare_preview_items(
enriched = dict(source)
if needs_structured_xlsx or needs_layout_raw:
enriched["raw_content"] = raw
sources[index] = enriched
continue
pages = extract_pdf_page_texts(raw)
extracted_text = "\n\n".join(page.text for page in pages if page.text)
if extracted_text != str(source.get("content") or ""):
logger.warning(
"跳过PDF文档噪声检测存储偏移量不一致 source_id=%s",
source["id"],
)
continue
enriched["document_noise_spans"] = detect_pdf_document_noise(pages)
if needs_pdf_noise:
# layout_hybrid 路径下也跑 PDF 文本规则噪声检测,
# 弥补 docling layout 模型对中文 PDF 页眉/页脚识别率低的问题。
pages = extract_pdf_page_texts(raw)
extracted_text = "\n\n".join(page.text for page in pages if page.text)
if extracted_text != str(source.get("content") or ""):
logger.warning(
"跳过PDF文档噪声检测存储偏移量不一致 source_id=%s",
source["id"],
)
else:
enriched["document_noise_spans"] = detect_pdf_document_noise(pages)
sources[index] = enriched
items = _build_preview_items(task, sources)
if not items and source_file_ids is None and is_unstructured: